Anthropic software engineering interview questions

The Process

Five rounds, and the one that fails most people isn't technical

01

30 minVideo call

Recruiter screen

Not a formality, and people do get cut here. Anthropic recruiters probe mission alignment on this call, then test it again properly at the end of the loop.

Turn up with an actual view on the work.
Enthusiasm about AI in general reads as thin. They want to hear what you think is hard, and where you think the field is wrong.
Ask about the org, not the perks.
The track you are routed to decides whether you sit the take home at all, so this is where you find out which loop you are actually facing.
Give one number for total compensation.
Let the recruiter decide how it splits between base and equity. Anthropic weights cash into base, so the base number is the larger half.

02

60 minCoderPad

Technical screen

Live coding in a shared editor, typically CoderPad, run by an engineer. One medium to hard problem in the language of your choice, and no AI assistance of any kind.

Pick a language you are fluent in.
There is one problem and it is medium to hard. Fighting your own syntax is the cheapest way to lose this round.
Say what you are optimising for.
Writing and reasoning quality is scored across roles here, and that includes how you explain a trade off out loud while you type.
Do not reach for a model.
Anthropic publishes a candidate AI policy at anthropic.com/candidate-ai-guidance, and every live round is covered by it.

03

2 hrs of work48 hour windowNo Claude allowed

Take home evaluation

Research and performance engineering tracks only; product and software candidates usually skip it. Two hours of working time inside a 48 hour window on a Python simulator of a fake accelerator, rewritten whenever a new Claude solves the current version.

Build your own debugging apparatus.
It ships with no visualisation tools, and part of the score is whether you write the inspector before you write the optimiser.
Expect a puzzle, not a benchmark.
Version three dropped realism for Zachtronics style puzzles on a tiny instruction set, with manual memory, VLIW packing, SIMD and multicore.
Budget the two hours deliberately.
The window is 48 hours but the working time is two. Running out is the single most common way this evaluation is lost.

04

4 to 6 roundsOne dayVirtual

The onsite loop

Four to six rounds of 60 minutes: a classic data structures round, an applied systems round, and system design covering distributed systems and inference serving. Research candidates also sit a heavily weighted ML depth round.

Concurrency turns up more than once.
Threading and synchronisation come up across rounds, not only in design. Go in comfortable with both.
Narrate while you work.
Verbal reasoning is scored next to the code. Silence gets written up as uncertainty even when the solution is right.
Leave the tests and the naming alone.
Readable and covered outranks compact. This is the closest thing Anthropic has to a house style.

05

60 minThen 1 to 2 weeks

Values and safety round

Every candidate sits it regardless of track, and it is weighted equally with the technical rounds. A concrete scenario, scored on whether you can articulate trade offs and hold calibrated uncertainty.

Read Dario before you sit it.
Machines of Loving Grace and The Adolescence of Technology are the common reference points, and it shows when you have not.
Take safety seriously without the doom.
Cheerleading and catastrophising both land badly. They want somebody who can hold the tension.
Have a position you can defend.
Agreeing with everything is the failure mode. Bring a view, argue it, and change your mind for a good reason.

Levels and Pay

Most of the offer is equity, and the equity is private

Pick a level to see how the package splits, what changes by city, and what the loop actually asks of you at that band.

San Francisco

New York

London

Dublin

Level	Experience	Total comp
Software EngineerEntry	1 to 3 yrs	~$420K
Senior EngineerSenior	3 to 6 yrs	~$565K
Staff EngineerStaff	6 to 10 yrs	~$785K
Principal EngineerPrincipal	10+ yrs	~$1.1M

Anthropic posts one US band. San Francisco, New York and Seattle appear on the same requisition at the same numbers, so the US tabs sit at parity by policy and not by rounding. Only the non US tabs change the figure, and most of what they change is the exchange rate.

Anthropic weights cash into base rather than running a bonus, so the base number here is the whole cash number.

Live Question Feed

What Anthropic is asking right now

40 questions · new ones added every 7 days

40 Verified

PRACTICAL

Implement a bank supporting deposits, transfers and scheduled payments, then extend it live when the interviewer adds a requirement you did not plan for.

Technical screen

60 min

CoderPad

Reported 9 times

Open

SYSTEM DESIGN

Design a multi region service with strict consistency requirements and walk through the failure modes.

Locked

VALUES

Describe a time your view on a technical direction changed, and what changed it.

Locked

CODING

Given a stream of events, maintain a rolling aggregate under a fixed memory budget.

Locked

DEEP DIVE

Walk through a system you owned end to end and defend the decisions inside it.

Locked

PRACTICAL

Extend a small service across several stages, each stage adding a new requirement.

Locked

SYSTEM DESIGN

Design a queue that preserves ordering per key while scaling horizontally.

Locked

CODING

Implement a cache with expiry, then make it safe under concurrent access.

Locked

VALUES

How would you weigh capability gains against deployment risk on a launch decision?

Locked

39 more questions available

The exact questions Anthropic is asking this week, tagged by round and difficulty.

Unlock All 40 Questions

Each question stays in the bank for 14 days, so everything here was asked in the last 14 days. Candidates who see them first pass at 3.2 times the normal rate.

Get Full Access

The Unfair Advantage

Why this actually works

01

Real questions, not guesses

Collected from engineers who sat the loop, then cross validated against other intel sources.

02

Fresh every week

The database updates every 7 days, so you never study stale problems.

03

Covers every round

Technical screens, the take home evaluation, systems design, ML depth and the values and safety round.

04

120+ companies tracked

Also interviewing at OpenAI, Google DeepMind or Meta? We have their questions too.

05

Difficulty and pattern data

Tagged by difficulty, round and category, so you know where prep time goes.

06

Candidate verified only

Reported by a candidate, then cross verified against multiple data points.

3.2×

Higher pass rate

120+

Companies tracked

3.2K+

Engineers using this

FAQ

Common questions

How are the questions sourced?

Directly from engineers who interviewed at Anthropic, then cross validated against other intel sources into a stack ranked list.

How often are questions updated?

We add new questions every 7 days and retire each one after 14 days, so the bank turns over twice in the window it covers. Anthropic rewrites the take home whenever a new Claude solves it, and the values round shifts as the research agenda moves.

Does Anthropic ask LeetCode questions?

Not really. The technical screen is one medium to hard problem in a shared editor, and the onsite pairs a classic data structures round with an applied systems problem.

Can I use AI tools in the interview?

Not in live rounds. Anthropic publishes explicit candidate guidance on this and has removed people from processes for ignoring it.

What is the values and safety round?

A scored conversation about deploying powerful AI responsibly, weighted equally with the technical rounds. Every candidate sits it, and it is the round that eliminates people who sail through FAANG loops.

What does Anthropic pay a Senior Engineer?

Around $565K in San Francisco, or roughly £350K in London. Forty percent of that is private equity, and there is no bonus at any level.

How long does the process take?

Three to six weeks, and the take home adds about a week. It only applies to research and performance engineering tracks, so most software candidates skip it.

Is this just for Anthropic?

No. GothamLoop covers 120+ companies including OpenAI, Google, Meta, Amazon and Apple.

The Anthropic interview cheat code nobody talks about

Five rounds, and the one that fails most people isn't technical

Recruiter screen

Technical screen

Take home evaluation

The onsite loop

Values and safety round

Most of the offer is equity, and the equity is private

What Anthropic is asking right now

Why this actually works

Real questions, not guesses

Fresh every week

Covers every round

120+ companies tracked

Difficulty and pattern data

Candidate verified only

Common questions

Pricing plan

Company Interview Guides